It is suggested that LoRA at r=32 with full vision encoder fine-tuning is a practical approach, reducing static peak VRAM from 36.2 to 10.8 GiB (parameters and optimizer states, activation memory excluded) without detectable performance loss.
Abstract
Deploying billion-parameter Vision-Language-Action (VLA) models on industrial hardware requires fine-tuning to bridge the embodiment gap. Full Fine-Tuning (FFT) provides maximal plasticity but requires data centre-grade GPUs. We present a systematic study of Low-Rank Adaptation (LoRA) for $\pi_0$, a flow-matching VLA, evaluated on four precision assembly tasks with a UR5e robotic manipulator. Across a sweep of LoRA ranks (r=8 to 256), allocation strategies, and component-freezing ablations, we find no statistically significant advantage of FFT over certain LoRA configurations. Performance saturates at r=32, and uniform allocation across the Vision-Language-Model (VLM) backbone and action expert proves sufficient. Freezing the VLM or restricting the vision encoder to LoRA significantly degrades performance, indicating that embodiment adaptation requires both semantic and visual plasticity. These results suggest that LoRA at r=32 with full vision encoder fine-tuning is a practical approach, reducing static peak VRAM from 36.2 to 10.8 GiB (parameters and optimizer states, activation memory excluded) without detectable performance loss.
NebulaVLA is presented, an asynchronous dual-frequency architecture that decouples high-level semantic reasoning from low-level action control, optimizing computational resources and modularity and introduces GESTURE-7, a unified language-grounded action representation.
Congyu Zhao, Shuai Tian, Xu Zhang et al.· 0 citations
State-of-the-art vision-language-action (VLA) models such as $\pi_{0.5}$ exhibit strong semantic understanding, instruction following and task behavior. However, when deployed on new robots, even minor mismatches in hardware configuration relative to pretraining can cause severe performance drops. Finetuning the VLA on in-domain expert data from the new embodiment improves performance on the expert task but leads to a loss in its original instruction following and behavioral priors. In this paper, we propose a self-supervised method that generates online interaction rollouts from the zero-shot VLA as additional training data for finetuning. Our experiments show this finetuning scheme yields strong multi-task policies that, on the target robot, (1) inherit prior tasks distilled from the zero-shot model, (2) enable generalist instruction following, while (3) learning new skills from expert data with improved sample efficiency. We demonstrate the success of our approach across test sets probing generalization on a real ALOHA robot and a new simulation benchmark in RoboTwin. Video results are available at https://self-supervised-control.pages.dev/
Prachi Garg, Steve Xing, Prahit Yaugand et al.· 0 citations
As Vision-Language-Action (VLA) models continue to scale in the number of parameters, the computational cost and resource requirements for domain-specific fine-tuning have become significant barriers to practical robotic deployment. While Parameter-Efficient Fine-Tuning (PEFT) techniques like LoRA (Low-Rank Adaptation) offer a potential solution, they often fail to match the task success rates of their fully fine-tuned counterparts. In this paper, we propose a novel hierarchical VLA architecture that achieves state-of-the-art performance while maintaining high parameter efficiency. Our model decomposes control into a high-level System 2 for instruction-conditioned semantic context encoding—a frozen PaliGemma-3B backbone with 0.12B trainable LoRA parameters and a reactive System 1 for multimodal fusion and action generation. To optimize training efficiency, System 2 takes only a single egocentric image, while System 1 recovers missing context by integrating wrist-view images via ResNet-34 and proprioceptive state history encoded with a single linear projection layer. This information is fused through a Transformer Encoder, and final action trajectories are refined via a Transformer-parameterized conditional flow-matching decoder. To improve task performance, we generate diverse candidates by sampling from N independently initialized Gaussian noise vectors and using different numbers of denoising steps K per sample, and then select the executed action using a Cal-QL (Calibrated Q-Learning)-based critic. Evaluated on the standardized LIBERO benchmark, our proposed model achieved a 98.1% average success rate, outperforming contemporary fully trained models across all task suites. These results demonstrate that strategic architectural design can enable parameter-efficient models to exceed the performance of full-scale fine-tuning, offering a viable path for high-performance robotics under constrained computational resources.
Woo-Kyoung Jeong, Yongwoo Gu, June-sup Yi et al.· 2026 23rd International Conf...· 0 citations
Similar to the natural capabilities of humans to sequentially learn new tasks, robots with Vision-Language-Action (VLA) models should possess lifelong learning ability to learn a new task when deployed in open-world environments. However, most recently proposed lifelong learning models aim to effectively learn the current task (plasticity) or maintain high accuracy on previous tasks (stability), while the plasticity-stability trade-off remains largely unsolved in robotic manipulation models. To address this fundamental challenge, we propose a cache-efficient lifelong Vision-Language-Action learning framework for robotic manipulation (i.e., LifelongVLA), which alleviates the plasticity-stability trade-off with a dual-timescale adaptation mechanism while achieving low-cost robotic deployment with a cache-efficient replay strategy. More concretely, we propose a dual-timescale LoRA gating module to decompose VLA adaptation into two lightweight pathways: a short-term adapter for plasticity and a long-term adapter for stable consolidation. These pathways are integrated via a task-aware gate, enabling explicit control of the plasticity-stability trade-off. In the skill replay phase, a cache-efficient stochastic replay strategy is proposed to preserve more balanced retention signals without full-trajectory storage. Finally, experiments show that LifelongVLA outperforms existing baselines, demonstrating efficient skill expansion, robust retention of learned manipulation behaviors, and reduced reliance on retraining for real-world deployment on an xArm robot.
Yao He, Gan Sun, Wenqi Liang et al.· arXiv.org· 1 citation
Xiao-Robotics-1 serves as a strong robot foundation policy that can be efficiently fine-tuned on complex, dexterous tasks with high data efficiency and across multiple simulation benchmarks, Xiaomi-Robotics-1 outperforms state-of-the-art methods.
Xiaomin Guo, Piao-Piao Jin, Jason Li et al.· arXiv.org· 16 citations· ⚡2
Large pretrained vision-language-action models achieve strong robot-manipulation performance, while compact alternatives have largely pursued efficiency by compressing the prevailing observation-to-action paradigm. We investigate whether predictive sensorimotor modeling can make more effective use of a limited parameter budget than direct observation-to-action mapping. We present PredVLA, a language-conditioned predictive-coding policy with only 0.68 million trainable network parameters and no robot-data pretraining. Its hierarchical recurrent dynamics predict visual features and proprioception, while observations influence latent state only through prediction-error-driven online inference. On LIBERO, PredVLA achieves an 86.9% mean success rate across the three short-horizon suites and 75.4% across all four suites. Under a controlled comparison using the same frozen front end, demonstrations, action decoder, and evaluation protocol, PredVLA achieves 3.7x and 7.4x the three-suite mean success rates of parameter-matched Transformer and LSTM behavior-cloning policies, respectively. A mechanism-by-mechanism transition to the recurrent behavior-cloning baseline shows that replacing the predictive pathway with direct observation input produces the largest single performance drop, accounting for approximately $70\%$ of the endpoint gap. Further ablations identify distinct contributions from training-time latent inference, test-time error regression, hierarchical timescales, and sensory prediction-error channels. Together, these results support predictive sensorimotor modeling as a strong inductive bias for compact language-conditioned robot control.
Hiroki Sawada, Shunichi Kasahara· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
